Pairwise Object Recognition

Pairwise geometric histograms (PGH) are a representation used for the recognition of rigid shape. They have been shown to be a robust solution for the recognition of arbitrary 2D shape in the presence of occlusion and scene clutter [1], [2], [5] and [6]. The method is both statistically founded and complete in the sense that a shape may be reconstructed from its PGH representation ([7]). The complete algorithm comprises a number of stages:

• Model image data (during training) and scene image data (during recognition) are processed by an edge extraction algorithm and the edge data approximated by line segments. (Treating individual edgels would be equivalent and line approximation is mainly done for speed).
• Model histograms (during training) and scene histograms (during recognition) are constructed for each line segment (reference line) by comparing the reference line to all other lines and making entries into a histogram according to the measured relative angles and perpendicular distances (see Figure 1). To account for errors in the measurement processes and to encode the variability in the way a shape may be segmented into lines, entries are blurred appropriately when being placed into the histogram. This representation encodes local shape geometry in a manner which is invariant to rotation and translation and is robust to missing data, line fragmentation and clutter.
• Scene line labelling is performed by finding good matches between scene histograms and model histograms using the Bhattacharyya metric [13].
• Object classifications are confirmed by finding consistent labelling within a scene using a probabilistic Hough transform, as described in [14].

   
Figure 1: PGH entry for a Single Line Comparison

Pairwise Parameters to be Optimised

The optimisation of the representation process can be viewed as a way of matching the stored model data to the difficulty of the recognition task. Thus we can expect to have to construct measures which quantify: repeatability, discriminability and compactness. As each of these cannot be measured using a single unified statistic we can expect to have to use optimisation procedures which operate on non-comensurate objectives. We will now define the motivation and construction of our measures.

Prior to constructing geometric histograms it is necessary to decide on the histogram scale to be used. The choice of maximum perpendicular distance , is driven by two conflicting requirements. On one hand should be small enough that the PGH represents local shape and is robust to missing data and occlusion. On the other hand should be large enough so that shape information in each PGH is distinct. Prior to the present work, rule-of-thumb used was to ensure that most of a shape was encoded into each histogram.

The simplest type of histogram is constructed by restricting angles to the range 0 to and distances to 0 to . This histogram is invariant to reflections of the shape data about the reference line and is described as mirror symmetric . Mirror reflection invariance is not always desirable and can be removed by using the direction of angles (clockwise or anti-clockwise) to extend the range of angle measurements to to . This doubles the number of entries in a histogram and thus the computation needed for histogram matching, but also increases the sparseness so improving robustness of matching in cluttered scenes. Depending on whether the point of intersection between the two line fragments (Figure  1 ) lies to the right or left of the reference line can be signed, thus again doubling the area of the histogram and extending the distance range to to .

Typically, there are three types of PGHs used (Figure  2(a) ): and , known as `mirrored', `rotated' and `directed' histograms respectively, which encode different degrees of geometric invariance . In order to add robustness to the PGHs a local circular window is applied centred around the mid-point of the reference line (Figure  1 ). The associated histogram is constructed by making entries for each line truncated to the circular window. The radius of this window is a continuous variable and equal to the maximum perpendicular distance ( ) value of the associated histogram. The parameter is normally set in accordance with the object size so that the maximum amount of useful data is constructed during matching. Its determination is a trade-off between the local scale of the object and the amount of clutter in the window. Associated with this parameter is the width of the bins on the perpendicular distance axis of the PGHs. The width parameter can be determined from the accuracy of the polygonal segmentation process. In this work we will simultaneously optimise the value of and the histogram type but the bin width remains at its theoretical value [7].

PGHs have been shown to be applicable to extremely large databases for hundreds of thousands of objects [6]. These results are based on an extrapolation of storage capacity from a much smaller test set. In order to construct a large recognition system automatically we will require an automated procedure for selecting the parametric representation of a line in terms of the free parameters specifying a PGH - the very nature of this problem requires us to take multiple competing objectives into account.